- In short
- A business problem is AI-shaped when solving it requires language, reasoning, or unstructured judgment rather than pure deterministic computation. Tasks with a single, formula-derivable correct answer belong in traditional code; ambiguous, judgment-heavy tasks with acceptable variance in phrasing are candidates for Claude. The first architecture decision is whether the problem needs a model at all.
The question that comes before every pattern
Before you choose an architecture, a model tier, or a context strategy, there is a prior decision that most designs skip: does this problem need a model at all? The Claude Certified Architect - Professional (CCAR-P) exam treats recognising the AI-shaped problem as a remember-level skill because everything downstream depends on it. Reach for Claude on a task that was really a lookup, and you have built an expensive, probabilistic solution to a problem that traditional code would have solved cheaply and exactly. The discipline is to interrogate the shape of the work before naming a tool for it.
A problem is AI-shaped when solving it requires language, reasoning, or unstructured judgment. Summarising a messy support thread, drafting a reply in a particular tone, classifying an ambiguous complaint, these turn on interpretation that has no single formula. A problem is not AI-shaped when it has one correct answer derivable from fixed rules: computing a discount from published criteria, looking up an account balance, validating a date format. Those belong in code that returns the same output every time.
- AI-shaped problem
- A business problem whose solution requires language, reasoning, or unstructured judgment rather than deterministic computation. AI-shaped tasks tolerate acceptable variance in phrasing and involve ambiguity; tasks with a single, formula-derivable correct answer are not AI-shaped and belong in traditional code.
Judgment and ambiguity are the signal
The reliable signal that a task is AI-shaped is the presence of genuine ambiguity that a human currently resolves with judgment. If two competent people could phrase the correct answer differently and both be right, the task tolerates the variance a model introduces, and that variance is not a defect. Reading intent out of a vaguely worded email, deciding which of five categories a novel complaint fits, synthesising three documents into one brief, none of these have a canonical string that a rule could produce.
The opposite signal is a single, checkable correct answer. If the task has one right output and any deviation is simply wrong, a model is the wrong tool because it offers no guarantee of producing that output every time. The fact that the work currently frustrates a human, or that its rules happen to be written in natural language somewhere, does not make it AI-shaped. A rule-based eligibility calculation is deterministic even when the policy describing it reads like prose.
The danger with a deterministic rule handed to a model is that it fails quietly on the messy inputs a demo never shows. Consider a threshold rule such as "any claim over £5,000 needs a senior reviewer." A model asked to enforce it routes clean figures correctly, but when the amount is buried in prose, "damages estimated around five thousand pounds," it treats the phrase as a loose estimate rather than a triggering value and lets the claim through. It has followed the letter of the instruction while drifting from its intent. A deterministic rule the business is counting on must be enforced by code that is right every time and can be tested and audited, not by a system that is right most of the time and throws no error when it is wrong.
Frustration is not a qualifier
A common trap in discovery is treating any painful, text-heavy task as an AI candidate because a person finds it tedious. Tedium is a signal that something should be automated, not that it should be automated with a model. A worker who hates manually copying figures between two systems has a real problem, but the fix is likely an integration or a script, not a Claude call that reasons about numbers it could simply be handed. Separating "this is slow and annoying" from "this needs interpretation" is the core of the skill.
What the CCAR-P exam trips candidates on
The exam tests two traps, and both are about over-applying the model. The first is assuming that any pain point involving text should be solved with an LLM, even when the underlying operation is a lookup or a calculation. A scenario will describe a frustrated worker and a natural-language process and invite you to reach for Claude, when the credited answer notices that the task has a single derivable answer and belongs in code.
The second is treating a well-specified, rule-based task as an AI problem because it currently frustrates a human. The trap is emotional: the person is suffering, so surely the impressive tool should help. The reliable reading holds the shape of the task separate from the mood of the person doing it. If a fixed rule produces the answer, the task is deterministic no matter how much anyone dislikes performing it by hand.
Worked example
A finance team complains that computing quarterly volume discounts is slow and error-prone. Each customer's discount follows a published tier table keyed to their order volume. A stakeholder asks you to 'add AI to fix the discount problem.' What do you recommend?
Start by naming the shape of the underlying task, not the pain. Computing a discount from a published tier table keyed to order volume is a deterministic calculation: for a given volume there is exactly one correct discount, and the rule is fixed. That is the signature of a task that belongs in traditional code, not a model call.
The frustration is real and worth solving, but it points to a missing piece of automation, not to an AI-shaped problem. A small function that reads the order volume and returns the tier discount is faster, cheaper, and, crucially, correct every single time, which is exactly the guarantee the finance team needs and a probabilistic model cannot offer.
The right recommendation is to solve the calculation in code and reserve Claude for any genuinely AI-shaped piece that discovery might surface alongside it, for example, drafting a plain-language explanation of the discount for a customer email. Notice how that second task turns on language and tone, tolerates variance, and has no single correct string, which is precisely what makes it AI-shaped while the calculation is not.
Common misreadings to avoid
Misconception
If a business process is described in natural language, it must be an AI task.
What's actually true
Misconception
A task that frustrates a human worker is a good candidate for a model.
What's actually true
How this shows up on the exam
Domain 1 questions on this knowledge point describe a business pain point and ask whether Claude belongs in the solution at all. The reliable reading is to classify the underlying task: language, reasoning, or unstructured judgment means AI-shaped; a single formula-derivable answer means deterministic code, regardless of how the process is described or how much it frustrates someone today.
This is the entry point to the whole discovery sequence. Once you have confirmed a task is AI-shaped, the next move is distinguishing live state from static knowledge in the ask, then mapping the pain point to a specific capability. The same instinct returns later when you recognise work that does not belong with Claude during decomposition, and it underpins every later choice on the autonomy spectrum.
A stakeholder says their team wastes hours each week manually checking whether submitted expense claims fall within fixed per-category limits, and asks you to 'use AI to automate the check.' The limits are published and never change mid-period. What is the best first response?
People also ask
How do you tell if a business problem needs an LLM at all?
What makes a task AI-shaped instead of deterministic?
Why is deciding whether to use a model the first architecture decision?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.